Commands & Config

Previous Next

Authentication commands

# Create a hub account
atria sign_up --username alice --email alice@example.com --password Secret123

# Authenticate (stores credentials in OS keyring)
atria sign_in --email alice@example.com --password Secret123

# Remove stored credentials
atria sign_out

sign_in uses email (not username). Credentials are stored via the OS keyring and reused automatically for subsequent hub calls.

Dataset commands

prepare_and_upload

Downloads a registered dataset, caches it to Delta Lake format, and pushes it to the hub as a versioned artifact:

atria datasets prepare_and_upload \
    --name cifar10/standard \
    --target_name my-org/cifar10-custom \
    --data_dir /data/cache \
    --is_public False

Key parameters:

Parameter Description
name Registered dataset name (e.g. "cifar10/standard")
target_name Hub artifact name; defaults to the dataset name if omitted
branch Hub branch to push to (default: "main")
data_dir Local directory for cached storage
is_public Whether the artifact is publicly visible
overwrite_existing Overwrite if a snapshot already exists on this branch

Internally: load_dataset_config(name) → config.build(...) → dataset.upload_to_hub(name=target_name, ...)

download

Pulls a dataset artifact from the hub:

atria datasets download --name my-org/cifar10-custom --branch main --download_dir ./data/cifar10/

Model commands

upload

Loads a model snapshot from a local snapshot directory (containing model.safetensors and metadata.yaml) and uploads it to the hub:

atria models upload \
    --name my-org/resnet50-cifar10 \
    --snapshot_dir ./runs/exp1/snapshot/ \
    --branch main

Key parameters:

Parameter Description
name Hub artifact name (e.g. "my-org/resnet50-cifar10")
snapshot_dir Path to local snapshot directory from pipeline.save_to_disk()
branch Hub branch to push to (default: "main")
is_public Whether the artifact is publicly visible

download

Pulls a model snapshot from the hub:

atria models download --name my-org/resnet50-cifar10 --branch main --download_dir ./models/

Why Fire over Click/Typer

Python Fire was chosen because: - CLI commands are plain functions — no decorator overhead, no schema definition - Fire automatically handles nested dicts and lists as CLI arguments - Adding a new command means adding a new function, not a new decorator chain

The trade-off is less control over help text and argument validation, but for a research tool where the Python API carries the complexity, this is acceptable.

Overview Usage